This study examines how a third-party agent can enhance reflection within dyadic collaborative learning by generating concept maps that offer feedback from an alternative viewpoint. Specifically, we compare an exocentric agent, which actively incorporates learner’s knowledge representations, with an egocentric agent, which remains relatively detached. This distinction aligns with the concept of erroneous examples, where structured knowledge—potentially containing inaccuracies or contradictions—encourages learners to critically evaluate and refine their conceptual understanding. To investigate this hypothesis, we developed a computational model of two teachable agents using the cognitive architecture ACT-R. The exocentric agent integrates the learner’s concept maps into its knowledge base, providing feedback that partly aligns with the learner’s understanding while introducing intentional errors or alternative links. In contrast, the egocentric agent generates feedback independently of the learner’s input, often producing less relevant concept maps. We conducted a controlled experiment with undergraduates, who worked in dyads to create and refine concept maps with the support of these agents. Results show that the exocentric agent significantly enhances learning. These findings underscore the importance of balancing difference and similarity in third-party agent feedback to optimize learning outcomes.